Abacus.AI

Initially known as RealityEngines.AI,[3] Abacus.AI is the world's first end-to-end AI and ML platform.[4] Abacus.AI can be used to set up data pipelines, specify custom machine learning specific transformations, train models and deploy and monitor them.[5] In addition to the core platform that can be used by data scientists to operationalize all types of machine learning models, Abacus.AI specializes in several use-case specific workflows including personalization, forecasting, and anomaly detection.[4]

The platform can be used to build enterprise scale deep learning systems and the company has invented several neural architecture search methods that can create custom neural networks from datasets based on a specific use-case. The company has over 10,000 users on the platform who have created over 30,000 models to date.[2]

Abacus.AI uses several neural architecture search techniques to create the right custom neural network based on the dataset and use-case. Once a data set is fed into the platform, it is evaluated. After evaluating the data set, the company decides the type of neural network architecture (NAS) that works best for a certain use case through a NAS algorithm.[3]

In addition to NAS, Abacus.AI has worked on ML interpretability and provides explanations for models created on it’s platform. In a research paper released on XAI, Abacus.AI has released the workflow associated with XAI-BENCH, a battery of synthetic datasets for benchmarking popular feature attribution algorithms. The dataset can be configured and re-engineered in order to simulate data from real-world using popular explainability techniques from several evaluation metrics.

In addition to the core algorithmic technology, Abacus.AI has an end-to-end deep learning operations platform that includes modules to stream data at scale, create a real-time feature store, and deploy a vector matching engine for vector similarity search.

Abacus.AI uses an autonomous AI generation service. This allows smaller companies to bridge the “data” gap with larger companies like Google or Facebook. While this would traditionally lead to the issue that the neural networks used are not being trained with data that is relevant to your business, Abacus.AI uses a machine learning technique that solves this. As per the company’s press release (Jan 28, 2020): “This technique creates synthetic data that augments the original dataset, and then trains a deep learning model on the combined dataset. These models yield up to 15% improvement in accuracy compared to models that are trained without using this augmentation technique.[6]”

This autonomous AI service emerged from two general pillars: generative adversarial networks (GAN) and network architecture search (NAS). Additionally, Abacus.AI has used “DAGAN”, GAN that is used for data augmentation, making synthetic data sets when not enough data is available to be able to train a neural network. They achieved this by using NAS “which finds the best architecture for GAN by trying various combinations of “cells”, basic primitives composed of neural network modules. Research and methods of this development were spearheaded by the companies’ leading researchers Colin White and Yash Savani. [Source]

Abacus.AI has a data wrangling module that connects to various data sources including S3,  GCP and Azure which makes it easy to set up data transformations for machine learning.

Once the data has been transformed, users can use one of the several use-cases supported by Abacus.AI to train a model using a Neural Architecture Search (NAS) technique known as Bayesian Optimization with Neural Architectures for Neural Architecture Search(BANANAS).[1]

Based on specific use-cases like demand forecasting, churn reduction, and name entity recognition, different NAS techniques will be applied to train the best model based on the size and the shape of the dataset and the use-case. For incident detection, Abacus.AI uses variational encoders.[7] For anomaly detections and account takeovers, they use unsupervised learning tools. [Source]

Additionally, the platform has a real-time deep learning system for enterprises. The service was originally confined to the automatic models that Abacus would offer. However, the service was provided as of August 2021, taking any TensorFlow or Pytorch model and plugging it into “an end-to-end, real-time deep learning system.[4]” This system is composed of modules developed by Abacus.AI that provide for the ingestion of data in a streaming fashion, as well as vector matching.[8] This allows companies to stream real-time events like clickstream data, online purchases, social media interactions, and media views from websites and “internet of things” sensors. This data can then be used to train deep learning models for generating contextual predictions in real-time.[9]

The technology could also be applied in recommender services. Rather than receiving a static set of data that has been compiled for a user, the system can ingest within a certain interval the sets of data a user has “liked” and use that to create a recommendation. [Source]

The data is then attacked by NAS which helps in creating cutting-edge models which are refined by a GAN in order to augment noisy or sparse data with synthetic data which further enhances data modeling.[1]

Abacus.AI has published multiple research papers on NeurIPS. Synthetic Benchmarks for Scientific Research in Explainable Machine Learning[10] releases XAI-Bench, a suite of synthetic data sets. NAS-Bench-x11 and the Power of Learning Curves[11] explores singular value decomposition and noise modeling to create surrogate benchmarks and delves into the learning curve extrapolation framework.[12] How Powerful are Performance Predictors in Neural Architecture Search[13] is the first large-scale study of performance predictors with an analysis of 31 techniques. A Study on Encodings for Neural Architecture Search[14] proposed a way to encode a neural network architecture to be efficiently manipulated by a search algorithm. Intra-Processing Methods for Debiasing Neural Networks[15] laid the foundations for the platform’s debiasing services.

Other notable published papers include An Analysis of Super-Neut Heuristics in Weight-Sharing NAS[16] on TPAMI, Learning by Turning: Neural Architecture Aware Optimization[17] on ICML, BANANAS: Bayesian Optimization with Neural Architectures for Neural Architecture Search[18] on AAAI, and Exploring the Loss of Landscape in Neural Architecture Search[19] on UAI.[20]

These research findings served as the groundwork for the techniques currently utilized by the Abacus.AI platform. Additionally, they have used these techniques in the prestigious competition CVPR, Unseen Data in Neural Architecture Search where they placed 2nd.[20]

2019: The company raised $5.3 Million in Seed Funding round led by Eric Schmidt.[3][21]

2020: The company raised $13 Million in Series A Funding round led by Index Ventures.[3]

2020: The company raised $22 Million in Series B Funding round led by Coatue.

2021: The company raised $50 Million in Series C Funding round led by Tiger Global Management.[2]

The company has raised over $90 Million till date.[2]