LOADING...

MIXED PATTERN

RECOGNIZER

Synergy between AI and expert wherein AI efficiently learns about the system (data) from the Human Expert to dynamically improve model performance with minimum effort from the Expert.

GO TO OVERVIEW

AI DEMO

* This is a simulated demo
Cats40%
Dogs40%
Birds40%

ABOUT

TECHNOLOGY

AI is capable of processing large data sets to extract correlation and other insights from the data. Human expert with deep understanding of the system is able to identify insights that are relevant, or those that an AI cannot observe / identify, and those that require knowledge of the world. While AI can handle large dimensional data, an expert can handle data only in lower dimensions. Hence, there is a need to connect the AI and expert to exploit the synergy between them. Labelling huge volume of data by Human Expert is cumbersome & time consuming, so AI smartly & efficiently involve the Expert for very minimal labelling process. And after every interactions with the Expert, the AI tries to pick the Expert’s brain by learning from expert input and generalize. After certain iterations, the AI should be able to infer like the expert, and quickly detect the underlying observation characteristics. The initial Feedbacks that Expert will require will have more labels along with explanations, but eventually when the AI is able to pick up Expert’s brain, the feedback can only be in validating id the label suggested by AI is right or not.

BUSINESS

USE CASES

Fraud Detection

In scenarios where data are not effectively labelled for training due to class imbalance, evolving trends, updation of rules, AI need to effectively collaborate with Subject Matter Expert to extract the right fraudulent behaviours and to avoiding predicting false positive.

Cyber Security

AI can help businesses better analyze threats and respond to attacks and security incidents which is evolving over time with the help of expert in the loop. It could also help to automate more menial tasks previously carried out by stretched and sometimes under-skilled security teams.

ARCHITECTURE

Clustering & Confidence Predictor Module groups the data initially. Q&A module samples few observations and request inputs from the Expert to label them and also provide the associated features/ reasoning. The Label Predictor generalizes the expert input, by assigning labels to all records, to be used by the DL model. DL model updates the network parameters and predicts labels with certain confidence. Clustering & Confidence Predictor Module takes the output from DL model and the cycle continues. Validation module accesses the performance of the DL model by validating its predictions for few observations by the Expert. This cycle also continues until the optimal performance of DL model is met.

Architecture

Explanation of modules:

  • Label Predictor: For each observation, based on the “nearest” labeled observation and specific rules provided by the user, predict the labels.
  • Clustering and Confidence Predictor: Based on clustering algorithms and specific rules provided by the user, as well as prediction confidence near each labeled centroid, create clusters and confidence levels at various distances from the clusters.
  • Q&A Module: Figures out what questions to ask, based on confidence levels, “holes” in the data etc. When it receives a response it updates the labels by calling label predictor
  • DL Model: Predict the labels provided by the user, or use the labels provided by the user as intermediate features that predict some eventual label (provided with the data)
  • Expert Feedback: can be about an observation (tag or label), or the why (binning logic, key features, etc.), or similarity, or correct/not correct.
  • Validation module: Measures performance of DL model by validating model predictions by the expert.

RAZORTHINK AI

DL MODEL

DESIGNER

Build this in Big Brain in just 4 steps within 30 minutes:

1Click on DL Model Designer and Drag & Drop Layers/ Operations
2Configure the parameters and so and so..
3Validate the model architecture
4Train it for your datasets.
Builder Parameters

    STATS

    200+PARAMETERS
    7 HOURSTIME TO TRAIN
    70,000 IMAGESTRAINING DATA
    Tensorflow logo
    DESIGN TIME IN VANILLA TENSORFLOW4 DAYS
    Razorthink Logo
    DESIGN TIME IN RAZORTHINK AI2 HOURS