Deep SYBIL: The Deep Learning Forecaster

chevron-icon
Back
project-presentation-img
Kevin R. C.
Project Owner

Deep SYBIL: The Deep Learning Forecaster

Funding Requested

$96,000 USD

Expert Review
Star Filled Image Star Filled Image Star Filled Image Star Filled Image Star Filled Image 0
Community
Star Filled Image Star Filled Image Star Filled Image Star Filled Image Star Filled Image 4.4 (5)

Overview

Deep SIBYL (D-SYBIL) is an advanced forecasting service and time-series research tool. It comprises a stacked ensemble of cutting-edge deep learning (DL) models that capture more complex, nonlinear, and multivariate patterns in time-series data. Immediate use cases of D-SYBIL range from mapping extreme weather patterns (like flooding) to tracking crypto price movements (like boom/bust cycles). D-SYBIL is an independent service from SYBIL, a DF-R2 awarded service that contains mainly statistical and machine learning (ML) base models. Nevertheless, D-SYBIL can combine with SYBIL to become the ultimate forecasting service on the SNET platform, covering a wide array of time-series applications.

Proposal Description

Company Name (if applicable)

Temporai

How our project will contribute to the growth of the decentralized AI platform

D-SYBIL contributes to the growth in the following ways:

  1. It provides a solution in a high-impact, domain-agnostic field of time-series forecasting. See the core 'problem' section below.
  2. It interoperates with other AI services like SYBIL (DF-R2) and MAUQ (DF-R3), forming a chain of AI services that is force multiplying.
  3. It provides a blueprint for effectively implementing projects involving deep learning/transformer models with sizable GPU computing, such as in other areas like NLP and GenAI.

The core problem we are aiming to solve

As the original SYBIL proposal states, the core 'problem' we are trying to solve is reliable and automatic forecasting for the masses. Forecasting is an essential yet challenging feat, especially without the proper resources or technical know-how.

SYBIL provides an ensemble "tried-and-true" statistical and ML forecasting model. Based on our testing, these models can work for more straightforward use cases (e.g., demand forecasting) with strong temporal characteristics (i.e., trend and seasonality). Still, they may come up short with more complex "real-world" scenarios such as:

  • complex nonlinear relationships between different factors (e.g., crypto prices and macro-events)
  • long-range temporal dependencies from much earlier time steps (e.g., events that happened years ago)
  • irregular patterns such as outliers, sudden spikes or volatility, and regime shifts (e.g., high-to-low interest rate environments)

Modeling these messier scenarios requires more flexible and sophisticated models, such as DL with neural networks, which D-SYBIL provides.

Our specific solution to this problem

D-SYBIL provides a forecasting service that can automatically run a combination (specifically ensemble stacking) of various deep-learning models, including:

  • RNN
  • Block RNN
  • D-Linear
  • N-Linear
  • N-BEATS
  • N-HiTS
  • TCN
  • Transformer-based

See Project Details for an explanation of each model.

Given the users' time-series data, D-SYBIL first trains these DL models and aggregates them with a meta-model. Then, it uses these trained model aggregates to forecast future time-series values. D-SYBIL can also be linked to SYBIL, so users can indirectly utilize D-SYBIL just by executing the SYBIL service itself. This is because SYBIL can get D-SYBIL's DL base model results or its aggregate results after triggering the D-SYBIL service, just like how SYBIL triggers the Onboard NeuralProphet service (awarded in DF-R2) to get NeuralProphet's hybrid model's result. That way, forecasting users can leverage a multitude of statistical, deep learning, and hybrid models all at their fingertips by triggering a single root service: SYBIL. The combined forecasts can capture both the traditional temporal features (like trend, seasonality, and autoregression) with the statistical models and more complex, nonlinear features with the DL models. 

See Project Details for further details on how D-SYBIL interoperates with SYBIL.

Project details

SYBIL Recap

Before detailing D-SYBIL, I will first summarize the original SYBIL: The General-Purpose Forecaster.

Architecture

SIBYL’s service uses stacked ensemble (or stacking) architecture. It trains a mix of statistical and ML models as base models; then their outputs are aggregated by a meta-model, which produces the composite forecast. Currently, SYBIL also includes one DL model RNN/LSTM, but that model will be transferred over to the D-SYBIL (in Milestones 5-6). That way, SYBIL does not have to utilize the more storage-intensive DL library PyTorch and only needs lighter statistical toolkits like scikit-learn and statsmodel.

API Schema

SIBYL has one service with two primary API functions for users to call:

  1. Train: users input training dataset for SIBYL to train with its stacking architecture, and outputs model parameters back to the user in JSON or another compatible file format.
  2. Forecast: user inputs the training dataset and model parameters to SIBYL, which then outputs the model forecast back to the user.

D-SIBYL will follow the same Train and Forecast API Schema as SYBIL, but may have different sub-parameter values as D-SIBYL may customize its DL base models differently than SIBYL.

Base Models

Here are the descriptions for each of the statistical and ML base models in SYBIL:

  1. Naive: forecast the value equal to the last observation
  2. Seasonal Naive: forecast the value equal to the last observation of the same season 
    • Naive and Seasonal Naive are SYBIL's baseline models, used to benchmark the more sophisticated models below
  3. Theta: decomposes the time-series data into Theta lines, which then estimate the curvature or trajectory of the time-series
  4. ETS: an exponential smoothing-based model that decomposes the time-series into Error (E), Trend (T), and Seasonal (S) components
  5. ARIMA: a combination of Autoregressive (AR) and Moving Average (MA) models with differencing (I) 
  6. TBATS: a combinatorial model that includes Trigonometric seasonality (T), Box-Cox Transformation (B), ARMA Error (A), Trend Components (T), and Seasonal Components (S)
  7. Linear Regression: fits a linear relationship between the explanatory (X) variables and dependent (y) variable by tuning the X coefficients to minimize the residual
  8. LightGBM: a gradient-boosted model known for being memory and computationally lightweight
  9. NeuralProphet: a hybrid statistical and DL model that contains trend, seasonlity, event/holidays effects, autoregression and lagged regression components, where the autoregression and lagged regression components utilize the custom fully connected network (FCN) called AR-NET
    • accessible through the Onboard NeuralProphet service (DF-R2) that SYBIL calls to 
  10. RNN: Recurrent Neural Network - type of neural network with memory cells of previous inputs, suitable for sequential data like such time-series or text data for Natural Language Processing (NLP)
    • will transfer to D-SYBIL in Milestone 6
    • SYBIL will add a couple of additional statistical models in D-SYBIL's Milestone 5, such as:
      • Kalman Filter, Fast Fourier Transform (FFT), or Wavelet

Meta-Models

SYBIL comprises of several meta-models that ensembles the base model results. They include:

  1. Naive Ensemble: a simple average of all the base model outputs, all equal weight
  2. Weighted Average: weight higher performing base models outputs more heavily based on the results on the validation or holdout set
  3. Linear Regression: determine base model sets through a regression analysis

Current Status

Temporai has completed 7 out of 8 development milestones for SYBIL. We are working on the last milestone to test, make final tweaks to SYBIL, and write the final report. SYBIL is also available on the SNET marketplace as well as through the SNET-CLI, using the CLI command: snet service print-metadata temporai sybil. We are also actively working with the SNET developers to get SYBIL's demo up and running in the marketplace.

 

D-SYBIL

D-SYBIL follows the same stacking architecture and API Schedule as SYBIL in order for clean interoperability between the two services. It's main distinguishing factor is its DL base models.

Base Models

Here are the descriptions for each of the deep learning base models in D-SYBIL:

  1. RNN: Recurrent Neural Network - type of neural network with memory cells of previous inputs, suitable for sequential data like such time-series or text data for Natural Language Processing (NLP)
  2. Block RNN: encoder-decoder neural network model that uses RNN as encoder and fully connected network (FCN) as decoder to produce fixed-length outputs
  3. D-Linear: lightweight model that first decomposes input data into trend and seasonality components using moving average (MA), then performs a linear transformation on each component to extract the features and produce forecasts
  4. N-Linear: lightweight model that first normalizes input data by subtracting the last value of sequence, then performs linear transformation to extract the features and produce forecasts
    1. D-Linear and N-Linear are D-SYBIL's baseline models due to them being relatively computationally lightweight compared to the other DL models, just like how Naive and Seasonal Naive are the baseline models for SYBIL
  5. N-BEATS: Neural Basis Expansion Analysis for Time Series Forecasting - comprises blocks of fully connected networks (FCNs) each producing its own output, the block outputs are combined to make the composite forecast
  6. N-HiTS: Neural Hierarchical Interpolation for Time Series Forecasting - suited to capture local and global patterns of various multivariate time-series data in a hierarchical format, employs block-based architecture like in N-BEATS 
  7. TCN: Temporal Convolutional Network - treats multivariate time-series as images, using convolutional neural networks (CNNs) instead of RNNs to capture temporal effects
  8. Transformer-based: time-series model that leverages the Transformer architecture (from NLP) and Attention Mechanisms to capture long-range dependencies and nonlinear relationships. Candidate Transformer-based models include:
    • TSMixer: Time Series Mixer
    • TFT: Temporal Fusion Transformer
    • TiDE: Time-series Dense Encoder

These base models are not set in stone and may be subject to change during the D-SYBIL implementation. For example, we found a more suitable DL base model and included that in the ensemble instead of the one listed here. We will also be using the darts Python library model to implement the DL base models, but will also explore alternative time-series libraries, including Google's temporian and sktime.

Meta-Models

In addition to the meta-models from SYBIL, D-SYBIL will also include a neural network-based meta-model. The NN model updates its weights based on the base model outputs, just like for the coefficients of the linear regression meta-model.

Integration with SYBIL service

As SYBIL is intended to be the primary general-purpose forecaster, SYBIL can trigger D-SYBIL service for its specialized DL model forecasts just as it triggers the Onboard NeuralProphet service for its NeuralProphet hybrid model. It is a unidirectional link since D-SYBIL will not call SYBIL for its statistical model as D-SYBIL specializes in DL, including PyTorch library. SYBIL can call D-SYBIL for the following two ways:

  • Base model level: SYBIL can get each of the D-SYBIL DL base model outputs and treats them and its own base models, then aggregates all statistical, ML, and DL base models with SYBIL's own meta-model 
  • Service (meta-model) level: SYBIL treats D-SYBIL service as a singular base model, as it has done for Onboard NeuralProphet service 

Integration with MAUQ service

Since D-SYBIL is an independent service, it can directly trigger MAUQ to provide D-SIBYL with probabilistic forecasts. Since many of D-SYBIL's DL models can already provide native probabilistic outputs, MAUQ can further calibrate their confidence interval on holdout labeled data using an uncertainty quantification technique (UQ) called conformal prediction

No integration with Onboard NeuralProphet service

D-SYBIL will not utilize Onboard NeuralProphet since NeuralProphet is a hybrid statistical and deep learning model, while D-SYBIL is strictly deep learning-based. Instead, SYBIL will trigger Onboard NeuralProphet alongside D-SIBYL.

SYBIL vs. D-SYBIL Champion-Challenger approach

Instead of combining SYBIl and D-SIBYL together, another forecasting approach the user can experiment with is running SYBIL and D-SIBYL services separately on the same time-series and then comparing their results. SYBIL's statistical models are the champion or incumbent models, while D-SYBIL's DL models are the challengers. If SYBIL's forecast performs well enough, then there is little need also to utilize D-SYBIL as it is computationally extensive. However, if SYBIL's forecasts are underperforming (i.e., complex time-series), then the user can try out D-SIBYL to see if it improves the forecasts over SYBIL. In a further experiment, the user can first try only baseline models of the respective SYBIL and D-SYBIL before trying the remaining base models of the two services. These examples are what make SYBIL and D-SYBIL ample time-series research tools.

Data and Privacy Considerations

Like SYBIL, D-SYBIL is a stateless service that does not store your train or forecast data or fitted model parameters after you use it. The user decides what time-series data he/she will input into D-SYBIL. We highly recommend not using this service (or the SNET platform in general) if your data contains highly sensitive information and is subject to legal/regulatory risks.

The competition and our USPs

Yes

Describe how your solution distinguishes itself from other solutions (if exist) and how it will succeed in the market.

Our solution distinguishes itself from other solutions in the following ways:

  • Our proven track record: We have successfully deployed SYBIL onto the SNET marketplace and are continuously releasing new versions of the service. Additionally, we are testing and refining SYBIL with various time-series datasets. We will export our methodology from SYBIL to D-SYBIL.
  • Our deep team: We have nine team members, all with relevant experience in AI development. Our team is composed of data scientists, research scientists, and SWEs.
  • Our focus on interoperability: D-SIBYL is designed to be interoperable with end applications and other services, especially with SYBIL. As SIBYL grows as a service, so will D-SIBYL since SIBYL will pass its user traffic to D-SIBYL every time SIBYL triggers D-SIBYL.
  • Our focus on real-world use cases: D-SIBYL is designed to be general-purpose and domain-agnostic so that it can work in various use cases involving complex time-series data in real-world settings.

Our team

Our D-SYBIL team consists of the following:

  • Kevin R.C.: project owner, data scientist, ML research engineer, prior crypto/blockchain experience (bio here)
  • Daniel S: data scientist, ML research scientist
  • Joseph K.: SWE, AWS cloud infra, SNET deployment
  • Zhiqian W.: SWE, architecture
  • Biran Y.: data scientist - blockchain/crypto
  • Peter G.: data scientist, risk modeler
  • Parv B.: data science grad intern - DL
  • Sairam V.: data science grad intern - time-series
  • Chinmay D.: data science grad intern - product mgmt
View Team

What we still need besides budget?

Yes

Describe the resources you still need

Temporai is seeking a part-time growth engineer with a multifaceted skillset in engineering, product management, and business development. Roles include:

  • Onboarding users and providing support for different Temporai services like D-SYBIL
  • Liaise between external stakeholders and internal technical/modeling teams
  • Test and fine-tune services or create new features to enhance usability
  • Creating, implementing, and evaluating ideas or strategies for growth
  • Represent Temporai in external meetings with SNET and ecosystem

Brownie points out that you are highly passionate about the intersection of AI + crypto/Web3 and time-series forecasting. If you think this is you, please reach out to kevin@tempor.ai.

Existing resources we will leverage for this project

Yes

Description of existing resources

We will leverage open-sourced Python libraries, such as forecasting library darts and the DL library PyTorch.

Open Source Licensing

gnu

Describe license details and, if applicable, list any components that are not subject to this license.

We will eventually open-source the entire D-SYBIL source code under the GNU GPL - GNU General Public License, just like for SYBIL and MAUQ.

Links and references

Here are the resource links to our prior awarded services:

SYBIL Service (DF-R2)

  • Proposal: link
  • YouTube Overview: link
  • YouTube Demo: link
  • Architectural Design Report Doc: link
  • GitHub Public Code Repo: link
  • SNET Marketplace: link

Onboard NeuralProphet Service (DF-R2)

  • Proposal: link
  • GitHub Public Code Repo: link
  • SNET Marketplace: link

MAUQ Service (DF-R3)

  • Proposal: link
  • YouTube Overview: link
  • GitHub Public Code Repo: link

Additional videos

See the YouTube links in the Links and references section above.

AI services (New or Existing)

SYBIL General-Purpose Forecaster

How it will be used

As explained in the Project Details section SYBIL will utilize D-SYBIL for its DL base model forecasts not the only way around.

MAUQ Probabilistic Predictor

Type

New AI service

Purpose

MAUQ service converts any model point prediction into a probabilistic prediction including forecasting models from SYBIL or D-SYBIL. MAUQ contains a single function called Quantity Uncertainty. Its input and output are specified below. This service is awarded in DF-R3 and will be available on the SNET marketplace shortly.

AI inputs

User inputs a calibration dataset (CSV file) and config parameters (JSON or YAML file) for MAUQ to fit the uncertainty measures. Just like in SYBIL MAUQ's front-end will convert the file format into JSON and concatenate the config parameters before passing it into the MAUQ service API.

AI outputs

MAUQ outputs that dataset with the uncertainty measures back to the user in JSON and then converted into a compatible file format like CSV.

Proposal Video

Placeholder for Spotlight Day Pitch-presentations. Video's will be added by the DF team when available.

  • Total Milestones

    14

  • Total Budget

    $96,000 USD

  • Last Updated

    20 May 2024

Milestone 1 - API Calls & Hostings

Description

This milestone represents the required reservation of 25% of your total requested budget for API calls or hosting costs. Because it is required we have prefilled it for you and it cannot be removed or adapted.

Deliverables

You can use this amount for payment of API calls on our platform. Use it to call other services or use it as a marketing instrument to have other parties try out your service. Alternatively you can use it to pay for hosting and computing costs.

Budget

$24,000 USD

Milestone 2 - GPU Cost Overrun Buffer

Description

GPU computational costs buffer for training and inference for these deep learning models on top of the 25% API Calls & Hosting.

Deliverables

None. I will request access to this buffer budget when needed during the implementation of Deep SYBIL (D-SIBYL) to pay for the expected additional GPU costs beyond what is allocated in the API Calls & Hostings budget.

Budget

$6,000 USD

Milestone 3 - Contract Sign

Description

Finalize milestone details and sign contract with SNET

Deliverables

- Finalize milestone details (including allocation of API Calls budget across development milestones) - Sign contract with SNET - Assemble team and set up tooling for D-SYBIL development (including JIRA GitHub Databricks AWS Cloud etc.)

Budget

$2,000 USD

Milestone 4 - Design Report

Description

Finalize D-SYBIL architectural design report

Deliverables

Write-up D-SYBIL architectural design report detailing its API schema code and libraries deep learning time-series models stacked ensemble design evaluation metrics applications and integration with SYBIL and other existing services.

Budget

$8,000 USD

Milestone 5 - Statistical SYBIL

Description

Replace RNN with at least two statistical models in existing SYBIL service

Deliverables

- Remove RNN and PyTorch library (reduce library size) in existing SYBIL service - Add at least two statistical models in SYBIL in place of RNN such as Kalman Filter Fast Fourier Transform (FFT) or Wavelet

Budget

$5,000 USD

Milestone 6 - RNN

Description

Set up a skeleton version of D-SYBIL service with RNN-based models as the first base deep learning models

Deliverables

- Spin-up AWS cloud environment (i.e. EC2 instance or Lambda) - Set up API gateway in accordance to API schema test hardcoded API calls and responses - Add RNN (and variants like LSTM) and Block RNN models to D-SYBIL service - Create example notebooks to run D-SYBIL service with RNN-based models as test

Budget

$6,000 USD

Milestone 7 - D-Linear & N-Linear

Description

Add D-Linear and N-Linear models to D-SYBIL as base models

Deliverables

- Conduct further research into D-Linear and N-Linear models including their hyperparameters and computational load - Integrate D-Linear and N-Linear models to D-SYBIL service source code - Test D-Linear and N-Linear models using the example notebooks to run D-SYBIL service

Budget

$4,000 USD

Milestone 8 - N-BEATS & N-HiTS

Description

Add N-BEATS and N-HiTS models to D-SYBIL as base models

Deliverables

- Conduct further research into N-BEATS and N-HiTS models including their hyperparameters and computational load - Integrate N-BEATS and N-HiTS models to D-SYBIL service source code - Test N-BEATS and N-HiTS models using the example notebooks to run D-SYBIL service

Budget

$6,000 USD

Milestone 9 - TCN

Description

Add Temporal Convolutional Networks (TCN) models to D-SYBIL as a base model

Deliverables

- Conduct further research into TCN models including their hyperparameters computational load and uses in multivariate or cross-sectional forecasting - Integrate TCN models to D-SYBIL service source code - Test TCN models using the example notebooks to run D-SYBIL service

Budget

$4,000 USD

Milestone 10 - NN Meta-Model

Description

Add neural networks (NN) to D-SYBIL as a meta-model

Deliverables

- Conduct further research into using neural networks (NNs) as a meta-model within a stacked ensemble - Integrate NN meta-model to D-SYBIL service source code - Test NN meta-model using the example notebooks to run D-SYBIL service

Budget

$4,000 USD

Milestone 11 - Transformer

Description

Add at least one time-series transformer or transformer-like model to D-SYBIL as a base model

Deliverables

- Conduct further research into the feasibility of time-series transformer/transformer-like models including Time Series Mixer (TSMixer) Temporal Fusion Transformer (TFT) or Time-series Dense Encoder (TiDE) - Integrate at least one transformer/transformer-like model to D-SYBIL service source code - Test transformer/transformer-like model using the example notebooks to run D-SYBIL service

Budget

$6,000 USD

Milestone 12 - SYBIL & MAUQ Integration

Description

Integrate D-SYBIL service into the original SYBIL service with MAUQ for probabilistic forecasts

Deliverables

- Add D-SIBYL ensemble into SYBIL as a single base model - Add D-SYBIL's DL base models into SYBIL base model options (and SYBIL will pass all the DL base models to D-SYBIL) - Allow users to toggle between both options - Test D-SYBIL integration using the example notebooks to run SYBIL service itself - Test SYBIL output calibrated forecasts with MAUQ service

Budget

$6,000 USD

Milestone 13 - SNET Deployment

Description

Deploy D-SIBYL onto SNET marketplace platform

Deliverables

- Deploy D-SIBYL API endpoint using SNET CLI - Publish D-SIBYL service onto SNET marketplace - Develop basic front-end marketplace UI - Test end-to-end from SNET marketplace UI

Budget

$5,000 USD

Milestone 14 - Final Report & Promotion

Description

Finalize D-SIBYL service write up a final report and promote

Deliverables

- Make final integration tests and perform any adjustments/fixes of D-SIBYL - Write up a final report for D-SIBYL - Work on demoing outreach and promotion of D-SIBYL (and other interoperable services like SIBYL) with the assistance of SNET and the wider ecosystem to gain user traction including creating an initial growth strategy plan

Budget

$10,000 USD

Join the Discussion (1)

Sort by

1 Comment
  • 0
    commentator-avatar
    HenriqC
    May 14, 2024 | 3:56 PM

    You are one of pretty few entities who have launched real services on the marketplace. That's highly appreciated and adds credibility quite a bit , thanks! Did you identify any specific challenges/ways to overcome them that might be common and help other teams on their journey ?

Reviews & Rating

Sort by

5 ratings
  • 0
    user-icon
    TrucTrixie
    May 9, 2024 | 1:50 AM

    Overall

    5

    • Feasibility 5
    • Viability 4
    • Desirabilty 5
    • Usefulness 5
    Focus on strengths in competition

    Deep Learning is a great prediction tool, I highly recommend this recommendation because of its feasibility and professionalism. The issue of making a team good at its product to be able to survive in the high competition of other forecasting tools with the same goal will appear in the near future. To achieve this, we need to pay special attention to ease of use, high accuracy forecasts, reasonable prices...

    user-icon
    Kevin R. C.
    May 11, 2024 | 7:46 PM
    Project Owner

    Thank you for your review TrucTrixie! Yes we will keep in mind about the forecast performance and the usability of D-SYBIL.

  • 0
    user-icon
    Joseph Gastoni
    May 6, 2024 | 8:14 AM

    Overall

    4

    • Feasibility 4
    • Viability 4
    • Desirabilty 4
    • Usefulness 4
    It has a good concept but careful consideration

    D-SYBIL has a promising concept, but careful consideration needs to be given to computational resource requirements and interpretability of deep learning models.

    Feasibility:

    • Moderate: The project requires expertise in deep learning and time-series forecasting.
    • Availability of sufficient computational resources (GPUs) for training deep learning models is crucial.

    Viability:

    • Moderate: The market for time-series forecasting is growing, but competition exists.
    • User adoption depends on accuracy, ease of use, pricing strategy, and integration with existing workflows.

    Desirability:

    • High: A more accurate and versatile forecasting tool for complex time-series data can be desirable for various industries.
    • The potential to capture non-linear relationships and long-range dependencies is valuable.

    Usefulness:

    • High: The project has the potential to improve the accuracy and flexibility of time-series forecasting.
    • The modular design allows for integration with existing forecasting services on the SNET platform.

    Besides, the project should consider:

    • The focus on deep learning models complements the existing statistical models offered by SYBIL.
    • Addressing interpretability of deep learning model forecasts can be beneficial for user trust.

    Here are some strengths of this project:

    • Offers a more sophisticated approach to forecasting compared to existing SYBIL service.
    • Targets complex time-series data with non-linear relationships and long-range dependencies.
    • Integrates well with the existing SYBIL service on the SNET platform.

    Here are some challenges to address:

    • The computational cost of training and running deep learning models.
    • Ensuring the interpretability and explainability of the deep learning model forecasts.
    • Demonstrating the added value of D-SYBIL compared to existing forecasting solutions.

    By addressing these challenges and focusing on user-friendliness and interpretability, D-SYBIL can become a valuable tool for advanced time-series forecasting on the SNET platform. A clear pricing strategy and targeting specific use cases will also be important for achieving user adoption.

    user-icon
    Kevin R. C.
    May 11, 2024 | 3:27 PM
    Project Owner

    Hi Joseph,

    Thank you for your thorough review. We will heed your advice on the interpretability and explainability of DL models. It will not be as easy as statistical models because deep learning's strength is learning complex or "deeper" features within the time-series, which naturally cannot be as easily interpreted. We'll do what we can to make DL more interpretable given this challenge. Ultimately, it is up to the user whether to use DL models to improve performance (by learning "deeper" features) at the expense of explainability. We will educate users of SYBIL and D-SYBIL about this trade-off so they can make the best judgment on this.

    As for the cost of training, we have allocated $30,000 of $96,000 for API calls and hostings, including the allocation of GPU computing. There is always a risk for us to overrun our compute budgets. Some resolutions we do limit GPU costs include:

    1. limiting the size of the datasets we use to train and forecasts
    2. developing a hybrid, flexible cloud strategy: including using AWS instances, Databricks, and even exploring bare-metal private cloud

  • 0
    user-icon
    BlackCoffee
    May 5, 2024 | 1:32 AM

    Overall

    5

    • Feasibility 4
    • Viability 4
    • Desirabilty 5
    • Usefulness 5
    Duties of each member

    It seems that the team has the necessary human resources to implement this proposal. I was impressed with the profiles of most of the members. The next thing is to start working once the funding is granted. If I and the community could know more about the specific division of work for each member, it would be great. This is a way to further affirm Kevin R.C's reputation in the community.

    user-icon
    Kevin R. C.
    May 11, 2024 | 1:48 PM
    Project Owner

    Hi BlackCoffee appreciate your review. So myself and Daniel are data scientists and ML researchers who will be deep-diving the different DL models for D-SYBIL. Zhiqian is a SWE managing the architecture of the D-SYBIL service. Joe is SWE who will take care of the infrastructure and deployment of D-SYBIL on cloud and SNET platform. Sairam, Parv, and Chinmay are data scientist interns who will implement the DL models into the D-SYBIL service under our guidance. And Biran is a data scientist who will be testing and evaluating the DL models and D-SYBIL service under various test cases. I will also be the project manager for D-SIBYL. This division of work is not set in stone since we all wear multiple hats and can fulfill each others' roles. Nevertheless, I hope this clarifies the division of work for each of our team members.

     

    Cheers,

    ~Kevin

  • 0
    user-icon
    Viclex Ad
    May 4, 2024 | 8:18 AM

    Overall

    4

    • Feasibility 5
    • Viability 4
    • Desirabilty 4
    • Usefulness 4
    "Deep SIBYL: Time-Series Forecasting Revolution"

    1. Feasibility:
      • The proposal demonstrates a high level of technical feasibility, considering the integration of various deep learning models and meta-models into the D-SYBIL service.
      • The team composition includes members with relevant expertise in data science, AI, and product management, enhancing the feasibility of project execution.
      • Budget allocation and milestones are well-defined, indicating a structured approach to development.
      • Rating: 5/5
    2. Desirability:
      • The proposal addresses a significant need for reliable and automatic forecasting in complex time-series data, making it highly desirable for users seeking advanced forecasting solutions.
      • Integration with existing services like SYBIL and MAUQ adds value and enhances the overall desirability of the D-SYBIL service.
      • The inclusion of diverse deep learning models and meta-models caters to a wide range of use cases, further increasing desirability.
      • Rating: 5/5
    3. Usefulness:
      • D-SYBIL\'s capabilities, including deep learning base models and meta-models, significantly enhance its usefulness in solving complex forecasting challenges.
      • The proposal outlines specific use cases such as mapping extreme weather patterns and tracking crypto price movements, highlighting the practical applications of D-SYBIL.
      • Integration with other AI services and the ability to provide probabilistic forecasts add to its usefulness and versatility.
      • Rating: 5/5
    4. Viability

    Viability is a crucial aspect to consider for any project, especially in the context of AI and technology development. Let\'s evaluate the viability of the D-SYBIL project based on several key factors:

    1. Market Demand:
    • There is a growing demand for advanced forecasting solutions, especially in fields like finance, weather prediction, and supply chain management.
    • D-SYBIL\'s focus on deep learning models for time-series forecasting addresses this demand for more accurate and nuanced predictions.
        
    2. Competitive Landscape:
    • The proposal acknowledges existing solutions like SYBIL and other AI services, indicating an understanding of the competitive landscape.
    • D-SYBIL\'s unique selling points, such as its deep learning base models and interoperability with other services, position it well against competitors.
    3. Technical Feasibility:
    • The proposal outlines a detailed technical architecture, including the integration of deep learning models, meta-models, and API schemas.
    • The team\'s expertise and experience in data science and AI contribute to the project\'s technical feasibility.
    4. Financial viability:
    • The budget allocation and milestones suggest a well-planned financial strategy for the project.
    • Integration with the SNET marketplace and potential revenue streams from service usage contribute to its financial viability.
    5. Regulatory and Ethical Considerations:
    • The proposal does not explicitly mention regulatory or ethical considerations, which are crucial factors, especially in AI development.
    • Compliance with data privacy laws, ethical AI principles, and regulatory requirements is essential for long-term viability.

    user-icon
    Kevin R. C.
    May 6, 2024 | 12:59 PM
    Project Owner

    Hi Viclex Ad, thanks for your thorough review of Deep SYBIL. Regarding your last point, Regulatory and Ethical Considers, we added the additional passage in the Project Details section:

    Data and Privacy Considerations

    Like SYBIL, D-SYBIL is a stateless service that does not store your train or forecast data or fitted model parameters after you use it. The user decides what time-series data he/she will input into D-SYBIL. We highly recommend not using this service (or the SNET platform in general) if your data contains highly sensitive information and is subject to legal/regulatory risks.

    Hope that addresses your concerns.

  • 0
    user-icon
    Max1524
    May 3, 2024 | 12:14 AM

    Overall

    4

    • Feasibility 4
    • Viability 4
    • Desirabilty 4
    • Usefulness 4
    Reconsider funding amount?

    Kevin R.C is a familiar face with Deepfunding from last season. The author has more or less built up a reputation with the community. This proposal is built quite tightly and is highly convincing. I am quite satisfied with this proposal.
    What needs to be done now is to add the full Company Name and provide more representative photos of members to enhance transparency.
    I'm thinking is $96,000 a lot of money compared to what it takes to get Deep SYBIL?

    user-icon
    Kevin R. C.
    May 3, 2024 | 2:25 AM
    Project Owner

    Hi Max1524, appreciate your review and your kind words. It's been an honor being part of this fantastic community as a builder. :)

    To respond to your points, my Company Name is Temporai, which should be in the Overview tab. As for the team members, yes I am trying to have their accounts set up on deepfunding.ai, including filling out their profiles and pics, so that is in progress.

    I believe that the $96,000 budget is reasonable for the following reasons:

    1. Compensation for technical skills and expertise to design, research, implement, and test these sophisticated DL models, which is non-trivial. The bulk of the milestones are dedicated to implementing these DL models.

    2. High expected computational costs to develop and test these DL models. The budget includes an additional $6,000 GPU costs on top of $24,000 for the required 25% on API Calls & Hostings. Therefore $30,000 of the $96,000 budget (31.25%) is dedicated toward infra and compute costs.

    3. Additional development to interoperate Deep SYBIL with other services, including SYBIL and MAUQ. This includes refactoring the existing SYBIL service, such as removing the RNN model and adding additional statistical models in its place. You can see a few milestones dedicated to achieving meaningful interoperability.

    4. Other general expenses such as operations, project management, SNET deployment (a big one), report writing, and user outreach/promotion. We have become more efficient with our processes thanks to our prior experiences with SYBIL and others, but these expenses still add up and need to be factored in.

    Therefore, the $96,000 ask is fair in my opinion.

Summary

Overall Community

4.4

from 5 reviews
  • 5
    2
  • 4
    3
  • 3
    0
  • 2
    0
  • 1
    0

Feasibility

4.4

from 5 reviews

Viability

4

from 5 reviews

Desirabilty

4.4

from 5 reviews

Usefulness

4.4

from 5 reviews