November 11, 2020 | Written by: Lisa Amini and Pin-Yu Chen
Share this post:
Launched in 2018, the Rensselaer-IBM Artificial Intelligence Research Collaboration (AIRC) is a multi-year, multi-million dollar joint venture boasting dozens of ongoing projects in 2020-2021 involving more than 80 IBM and RPI researchers working to advance AI. The collaboration is part of the IBM AI Horizons Network (AIHN), a program dedicated to advancing the science of AI and enabling the use of AI and machine learning in research investigations, innovations and applications of joint interest to IBM and our university partners.
Neural networks are an effective way to determine patterns in data, but not all neural networks are created equal. The ultimate goal when designing a neural network is finding an optimal set of parameters for solving a particular task, meaning loss is minimized. As you might imagine, finding the right neural network model for a given problem is no small feat. Often, trade-offs must be made between achieving high accuracy, being robust to adversarial attacks, and determining the time and computational resources required to train an accurate and robust neural network. Discovering techniques to efficiently train neural networks that balance accuracy and robustness goals is a key focus area of the Rensselaer-IBM AIRC.
A team of AIRC researchers from these two organizations recently made a significant breakthrough in efficiently searching for new models that balance optimality and robustness. In the published paper “Optimizing Mode Connectivity via Neuron Alignment” — to be presented at the upcoming NeurIPS 2020 conference — the researchers propose a novel neuron alignment technique to address the permutation ambiguity in the space of neural network models. The result of their research is to effectively reduce the loss barrier between local optimal models and aid in finding more accurate and robust models. They showed that the proposed neuron alignment technique can efficiently find a model with improved robustness and accuracy, which is missed by existing methods.
This joint IBM-RPI project studying the loss landscape of deep learning models and how sets of locally optimized parameters are connected, is one of many successful ventures to emerge from the AIRC.
Another study presented at ECCV’20 in August, for example, addressed security threats to deep neural networks (DNN) that leave them vulnerable to tampering and could bias results. In the paper, “Practical Detection of Trojan Neural Networks: Data-Limited and Data-Free Cases”, IBM and RPI researchers describe how DNNs can be manipulated by an adversary known as the Trojan attack (or poisoning backdoor attack).
AI researchers often leverage pre-trained AI models to provide transfer learning that speeds up subsequent development. If a pre-trained model has been compromised by a Trojan attack, applications built using that model are at risk for tampering that biases machine learning systems and undermines trust in AI. The IBM and RPI researchers proposed a new technique to discern the fidelity of a pre-trained AI model. Their data-limited Trojan network detector works even when only a few data samples are available for TrojanNet detection. The research describes how an effective data-limited detector can be established by exploring connections between Trojan attack and prediction-evasion adversarial attacks.
AIRC researchers jointly filed for 11 patents and published 18 studies in 2020 at top AI conferences and journals, including this year’s NeurIPS, ICML, KDD, AAAI and ECCV. AIRC projects align well with IBM Research’s top priorities in trusting AI, advancing AI and scaling AI and are expected to have a large impact on businesses as AI adoption becomes more widespread.
- Trusting AI includes: Explainability, fairness, robustness, generalization, ethics, governance and regulation, causal inference
- Advancing AI includes: AI for scientific discovery, AI for cyber security, natural language processing, conversation system, foundational AI algorithms
- Scaling AI includes: Large-scale optimization for deep learning, AI system optimization, auto AI, secure and hybrid cloud + AI
Extern and Scholar Programs
In addition to producing high quality AI research, AIRC helps IBM and RPI cultivate much-needed AI talent by supporting long-term research projects in the areas of trusting AI, advancing AI and scaling AI. An especially beneficial element of this collaboration is its scholar and extern programs, where RPI graduate students – some of the top AI PhD students worldwide – receive in-depth research collaboration, mentorship, career development, as well as short- and long-term project planning through the AIRC.
Pioneering a new model of collaboration between industry and academia, the AIRC has a robust externship program. These are one-term fellowships aimed at accelerating the application of AI, machine learning, natural language processing (NLP) and related technologies, proposed by IBM researchers. Our extern program has hosted more than 30 projects since 2019; 25 of them over the past summer. With the scholar’s program, students are offered a three-year mentorship as well as two residency terms at IBM Research.
Looking ahead, AIRC projects will continue to advance new research that aims to address areas expected to have a major impact on AI’s future and its potential application for business. These projects include those aimed at using AI to accelerate scientific discovery, designing a fair collective decision-making AI method, studying sample complexity for achieving good generalization, devising novel approaches to mitigating domain shifts and efficient deep learning with less data, and using human-inspired learning principle to design efficient and generalizable decision-making AI systems.
Inventing What’s Next.
Stay up to date with the latest announcements, research, and events from IBM Research through our newsletter.