Ricky Ma

Hello! I am currently a machine learning engineer at Rebel Space Technologies building tech that enables autonomous, ad-hoc wireless connectivity in dynamic environments for mission-critical operations - come check us out!

In my free time, I am developing a platform to supercharge human rights and environmental law protection at Lautonomy. My co-founder and I hope to improve redress and access to legal representation for torture survivors and asylum seekers by removing the complexity that systemically blocks fair and equitable access to law. I also collaborate closely with Tobias Gerstenberg at Stanford’s Causality in Cognition Lab, researching computational models of counterfactual inference and how causal networks and simulations can be used to model these cognitive processes in humans.

I completed my BSc. at UBC studying Cognitive Systems with a focus in Computational Intelligence and Design, directed by the Computer Science Department. I’ve spent time at the WiNMoS Lab with Yanjie Dong doing research in decentralized reinforcement learning. I’ve also worked with Matthew Cooke at Snyder Lab on scientific software development and data analysis for behavioural neuroscience.

Previously, I developed a novel framework for unsupervised anomalous object detection at RecyclEye. At Algomo, I created an interactive model-understanding and evaluation platform for chatbot data and feedback, and invented a brand-new method for named entity recognition using OpenAI’s GPT-3.

Research interests:

Multi-agent deep reinforcement learning
Self-supervised and semi-supervised learning
Causal reasoning and its effects on interpretability and generalization
Causal learning and inference in neural networks and time-series models
ML and AI for social good

news

Jul 16, 2021	Excited to announce that I will be officially joining Lautonomy as CTO. I hope to use my knowledge and experience to contribute to Lautonomy’s goal of promoting equal access to legal services for all.
May 12, 2021	I am joining the Causality in Cognition Lab at Stanford University as a visiting student researcher! I will be studying computational models of counterfactual inference.
Apr 26, 2021	I’m graduating soon!

selected publications

arXiv

ODDObjects: A Framework for Multiclass Unsupervised Anomaly Detection on Masked Objects

Ma, Ricky

arXiv preprint 2021

Abs PDF Code

This paper presents a novel framework for unsupervised anomaly detection on masked objects called ODDObjects, which stands for Out-of-Distribution Detection on Objects. ODDObjects is designed to detect anomalies of various categories using unsupervised autoencoders trained on COCO-style datasets. The method utilizes autoencoder-based image reconstruction, where high reconstruction error indicates the possibility of an anomaly. The framework extends previous work on anomaly detection with autoencoders, comparing state-of-the-art models trained on object recognition datasets. Various model architectures were compared, and experimental results show that memory-augmented deep convolutional autoencoders perform the best at detecting out-of-distribution objects.
F1000

Pathfinder: open source software for analyzing spatial navigation search strategies

Cooke, Matthew B, O’Leary, Timothy P, Phelan, Harris, Ma, Ricky, Brown, Richard E, and Snyder, Jason

F1000Research 2020

Abs PDF Code

Spatial navigation is a universal behavior that varies depending on goals, experience and available sensory stimuli. Spatial navigational tasks are routinely used to study learning, memory and goal-directed behavior, in both animals and humans. One popular paradigm for testing spatial memory is the Morris water maze, where subjects learn the location of a hidden platform that offers escape from a pool of water. Researchers typically express learning as a function of the latency to escape, though this reveals little about the underlying navigational strategies. Recently, a number of studies have begun to classify water maze search strategies in order to clarify the precise spatial and mnemonic functions of different brain regions, and to identify which aspects of spatial memory are disrupted in disease models. However, despite their usefulness, strategy analyses have not been widely adopted due to the lack of software to automate analyses. To address this need we developed Pathfinder, an open source application for analyzing spatial navigation behaviors. In a representative dataset, we show that Pathfinder effectively characterizes the development of highly-specific spatial search strategies as male and female mice learn a standard spatial water maze. Pathfinder can read data files from commercially- and freely-available software packages, is optimized for classifying search strategies in water maze paradigms, and can also be used to analyze 2D navigation by other species, and in other tasks, as long as timestamped xy coordinates are available. Pathfinder is simple to use, can automatically determine pool and platform geometry, generates heat maps, analyzes navigation with respect to multiple goal locations, and can be updated to accommodate future developments in spatial behavioral analyses. Given these features, Pathfinder may be a useful tool for studying how navigational strategies are regulated by the environment, depend on specific neural circuits, and are altered by pathology.