Freddy Trinh

Excited to share that our latest research paper titled "Learning Chemotherapy Drug Action via Universal Physics-Informed Neural Networks (UPINNs)" by Lena Podina and Ali Ghodsi @Lena Podina, @Ali Ghodsi, and myself, is now available on arXiv! Quantitative systems pharmacology (QSP) plays an important role in evaluating drug effects and toxicity before clinical trials. However, constructing a QSP model requires significant manual literature distillation, parameter fitting, and simplifying assumptions. Our research showcases the potential of UPINNs in learning unknown terms in pharmacodynamic and pharmacokinetic models, paving the way for more efficient and accurate drug development processes. Read the full paper here: https://lnkd.in/eGtk9SQT #Research #Pharmacology #MachineLearning #NeuralNetworks #DrugDevelopment #QSP #UPINNs #AI #DeepLearning

57

New paper published: ETTrack: enhanced temporal motion predictor for multi-object tracking https://buff.ly/415mtSo

12

https://lnkd.in/gBXgQmeN Often for junior learners, to bridge two concepts, but yet acquire an understanding of pathophysiological mechanisms may imagine false causation, when correlation might be a better explanation for the observation. This might explain the hallucination effect aka imagining things that are not real also seen with AI. It is imperative that real humans invest in a strong foundation, that is not merely based on memorization, but supported and fostered by critical thinking and curiosity. This will empower one to distinguish oneself from the threats of "AI". Biases can sneak into AI systems through the creators' conscious or unconscious skewed data sets used for training. As custodians of human knowledge, prudent care is essential. Especially when machines have clearly become far by much more eloquent these days. Spotting errors in logic can get tough. #AIvsrealintelligence

12

I have written a brief accessible introduction to the latest paper Marta Andronic and I have written, NeuraLUT, on hiding neural network complexity in synthesisable Boolean functions. https://lnkd.in/e55MhuMV

47

I'm thrilled to share our research, "Generalized Deep Learning for Histopathology Image Classification Using Supervised Contrastive Learning," now published in the Journal of Advanced Research (Impact Factor: 11.4)! 🎉 Read the full paper here: https://lnkd.in/g8HXh_GT In this work, we introduce HistopathAI, a novel hybrid network that leverages supervised contrastive learning (SCL) and hybrid deep feature fusion (HDFF) to enhance the accuracy and generalizability of histopathology image classification. Key Highlights: 1. Demonstrated superior performance across seven public datasets and one private dataset, showcasing robustness in diverse histopathology domains. 2. Tackled challenges like imbalanced datasets, paving the way for improved diagnostic precision and efficiency in clinical pathology. 3. Aimed to support the global shift toward digital pathology, providing a powerful AI tool to improve cancer diagnosis and patient outcomes. I am grateful to my supervisors, Ewan Millar and Erik Meijering, for their invaluable guidance and collaboration on this project. #Research #DeepLearning #Histopathology #AIinMedicine

57

4 Comments

We have a new fully-funded PhD advertised on the CDT, "Better Personalization of Deep Learning-Enhanced Hearing Devices" See University of Sheffield list https://lnkd.in/e2CkWTRG Hearing loss affects over 5% of the world’s population, making it a major public health concern. Hearing aids are the most commonly prescribed treatment, but many users report they do not perform well for listening to speech in noisy situations. Breakthroughs in deep learning and low-power chip design are driving the next generation of hearing devices and wearables, with the potential to revolutionize speech understanding in challenging listening environments. For example, Apple’s AirPods Pro have gained FDA approval as hearing aids for mild to moderate hearing loss, and Phonak has introduced deep neural network-equipped devices that dynamically enhance speech clarity in noisy environments. However, training these approaches to work in general settings and to suit individual preferences remains a critical challenge. To improve deep learning-enhanced hearing aids, we require metrics that predict how well a given hearing aid algorithm will perform for a specific user in a particular acoustic environment. Existing approaches often rely on oversimplified assumptions about listener preferences, which are captured using basic metrics. For example, it is often assumed there is a well-defined target speaker and that processing should maximise noise suppression while preserving quality. These simple metrics do little to capture users’ needs in more complex settings, such as trying to engage in multiparty conversations in a busy restaurant. The project will explore a variety of methods to understand hearing device user preferences in more complex settings, including leveraging virtual reality (VR) to simulate diverse acoustic environments and hearing aid algorithms. VR offers the advantage of creating immersive and controlled scenarios where users can directly experience and evaluate different algorithmic configurations. This approach allows the systematic measurement of user preferences across a wide range of conditions, ensuring both ecological validity and experimental rigor. From this understanding new algorithm quality metrics will be derived for optimising existing deep-learning enhancement approaches in a more user-dependent manner. The project will be based at the University of Sheffield and co-supervised by experts from both Sheffield and the University of Salford, collaborators on the ongoing EPSRC-funded Clarity Project . The Clarity Project focuses on improving speech-in-noise understanding, making it a natural foundation for this work. The Royal National Institute for Deaf People (RNID) will act as a key partner, offering additional expertise and a crucial end-user perspective.

20

2 Comments

"In our tests, the next model update performs similarly to PhD students on challenging benchmark tasks in physics, chemistry, and biology. We also found that it excels in math and coding. In a qualifying exam for the International Mathematics Olympiad (IMO), GPT-4o correctly solved only 13% of problems, while the reasoning model scored 83%. Their coding abilities were evaluated in contests and reached the 89th percentile in Codeforces competitions. You can read more about this in our technical research post." It is no longer how well and how fast you can code, its how well you can drive AI to do it for you. I have been developing Apps in Android Studio for just over a week, quite complex apps. I am getting the process down pat for communicating with the AI and keeping it on track without losing track yourself of what the code is doing. I think the secret is UML and a specification to instruct a custom GPT to develop a specification which you load back into the custom GPT. Next I am going to teach my daughter how to implement this, she lives with a cognitive disability that really impacts her ability to understand maths and logic. She is still thinking about what she wants to develop, I am keen to see what she comes up with. I doubt very much she will feel like she has a disability once we are through. https://lnkd.in/gawtRJaU

5

2 Comments

New paper on the stability of neuronal representations in the retrosplenial cortex (RSC), done in collaboration Michael Goard at UCSB. Briefly, we trained mice to associate a visual context with a motor choice while recording neuronal activity in RSC across days. Using SVM classifiers, we decoded task variables in both, individual neurons and populations. We identified neurons with different levels of stability, with visual context and trial outcome displaying higher stability, and motor choice exhibiting a faster decay. In summary, our findings show an important characteristic of association areas, where diverse streams of information are stored with varying levels of stability, balancing representational reliability and flexibility according to behavioral demands. https://lnkd.in/dHbXSnxe

16

1 Comment

Fascinating research on the neural substrates underlying visual memories! A study on monkeys' primary visual cortex revealed unique neuronal responses to different memory contents, distinct from external visual stimuli. The findings suggest that the mind's eye encodes memories beyond typical visual processing areas. This study sheds light on the visual encoding of memories, offering insights into the power of our ability to recall vivid images in the absence of actual stimuli. Read more about this intriguing research: https://lnkd.in/d6PbM8BG

1

As we rapidly approach London data week and the national election, I have been asked several times a version of the question “what should the next government do to make AI work in the NHS?” Fortunately, (or unfortunately depending on your opinion of AI/me/academia), I spent three years attempting to answer that question. The very very very looooooooong answer I came up with is in my thesis “Designing an Algorithmically Enhanced NHS” which is available online here: https://lnkd.in/eA7tG2TA I am, however, aware that very few people have either the time or the desire to read c. 90,000 words about NHS information infrastructure. I did gave a fifteen minute overview of the findings and conclusions at the Nuffield Summit in March but that does lose (rather a lot) of the detail. SO in an attempt to find a balance between the too brief and the too long, I have recorded an 80 minute version that answers the burning question “How can we enable the NHS to capitalise on the benefits of Algorithmic Clinical Decision Support Software (AI) while proactively mitigating the ethical risks of harm and the practical risks of failure?” You can watch it on YouTube below (it has captions in case you prefer not to listen to my voice (can’t blame you)). https://lnkd.in/eG-NDr6M - The slides from the above recording are here: https://lnkd.in/ejNsq59n - The completion of my thesis (and so its video version) would not have been possible without the support of a small army of people, all of whom are acknowledged here. https://lnkd.in/eEizmwyd - I also talk a lot about what ‘policy’ does and does not cover in the video. The full thesis goes into detail about exactly which policy documents I am referring to. In the interest of not boring people to tears - the video does not. However, if you want the full list of the policy documents I base my analysis on then it is available here: https://lnkd.in/e3MRyTz4

105

9 Comments

It's Sunday and Sunday's are for weeknotes so here are this week's notes: https://lnkd.in/eVN9gzwc Featuring: ✍ Writing a paper for the WHO Bulletin on the ethics of the Rapid Digitisation of Healthcare (not my title) and trying to move the conversation beyond privacy, safety etc. 📑 Publishing our systematic review examining the real world effectiveness of population health risk stratification algorithms: https://lnkd.in/ekc8BgrQ with Christopher Oddy, Joe Zhang & Hutan Ashrafian 📽 Turning my 90,000 word thesis on what needs to be done to make AI work in the NHS into an 80 minute video/podcast (see below). 🏛 Organising two major AI in Healthcare events with Luciano Floridi, and Renée Sirbu https://lnkd.in/e2UHQ7k2 💻 Revising our (Kassandra Karpathakis, mine, and Luciano Floridi) paper on a Justifiable Investment in AI for Healthcare https://lnkd.in/eJcKSC3Y 🧠 and more things I worked on, did, thought about, and read.

9

2 Comments

Just finished the course “NumPy Essential Training: 1 Foundations of NumPy” by Terezija Semenski! Check it out: https://lnkd.in/d6nvqSHq #machinelearning #numpy #scientificcomputing.

3

New paper! We argue – with paper and demo, to be presented at FDG next month in Worcester, MA – that the most promising avenue for LLMs in interactive narrative is in addressing the natural language generation (NLG) problem of surface text realization. For this project, we don’t want to cede control over story progression, game state, character features, etc. to ML systems. Those are all explicitly maintained in a symbolic (logic-based) AI system. But we do want to generate, at each point in the game where dialogue might be needed, a contextually relevant line of dialogue. Existing procedural narrative systems often already turn to natural language generation (NLG) for that problem, e.g. template- or grammar-based systems. We use large language models (LLMs) instead, which we argue have advantages in quality of surface text and ability to integrate multiple contexts. But we still retain the authorial control typical of classical NLG pipelines, in the sense that we use the LLM solely for surface text production. https://lnkd.in/eAug_7DW

28

1 Comment

Finally out in the public domain, still sometime until they will officially be on pubmed, but nonetheless two great achievements for the Nalesso's lab. I am very proud of the work of the PhD students and post-doc involved in their making and I cannot thank enough my collaborators Mohit Kapoor and Ian Clark for believing in the Wnt project. Happy to add yet some more complexity to the regulation of this signalling pathway. These projects have been generously funded by a #MRC and #Academy of Medical Sciences. Two more papers coming soon... microRNA-544a as a new modulator of the Wnt-signalling network in the articular cartilage and osteoarthritis bioRxiv 2024.11.28.625937; doi: https://lnkd.in/e-yvEvmy CaMKII induces an autophagy-dependent anabolic response in Articular Chondrocytes https://lnkd.in/gYZMkiqX

27

Large Language Models (LLMs) are not truly capable of emergent abilities; they are simply very effective at following instructions. This is the key conclusion of a fantastic paper published last month. The authors further assert that this holds true regardless of the LLM's scale, effectively suggesting that increasing the size of models with the same architecture will not result in the emergence of reasoning or any other latent, potentially dangerous abilities. What does this mean for the use of LLMs in healthcare? It implies that using a zero-shot model for critical healthcare tasks can be very risky and more prone to hallucination. Another significant risk is the potential disconnect between the AI/ML/data science team and the deployment team in an enterprise setting, especially if the LLM is off-the-shelf. Since In-Context Learning (ICL) and prompt engineering are more likely to occur at the deployment stage, any lack of rigor in this step—whether intentional or unintentional—can lead to safety issues. This could happen if there is misplaced trust in the model’s ability to "reason," possibly due to overselling by the vendor. https://lnkd.in/gf3qQjfc

20

2 Comments

I am really happy to share that our paper ‘Common neural mechanisms supporting time judgements in humans and monkeys’ was published in PeerJ!  (link: 10.7717/peerj.18477). In this project, we wanted to assess whether the neural mechanisms supporting time estimation are similar in monkeys and humans. For that purpose, we compared event-related responses (ERPs) of one monkey and 24 humans while performing a temporal bisection task. In this task, participants had to determine whether the duration of a time interval was longer or shorter than a prototype. In line with previous work, we find that post-interval ERPs in humans are significantly different between stimuli that were classified as ‘long’ or ‘short’ (even the objective duration was the same). Interestingly, we find the same pattern of results in one monkey when we look at post-interval ERPs and firing rates in pre-SMA. Together, our results show that common neural mechanisms support time categorisation in monkeys and humans. I would like to thank my collaborators Saskia Haegens (Columbia University), Elie Rassi (Radboud University Nijmegen), Hugo Merchant (UNAM) and German Mendoza (UNAM).  

19

I'm excited to share that our new paper "Intrepid MCMC: Metropolis-Hastings with Exploration" just hit Arxiv. Link below 👇 https://lnkd.in/eyWKWGad This paper was really the brainchild of my PhD student Promit Chakroborty who rigorously proved (see the Appendices!!!) that standard random walk Metropolis-Hastings (MH) algorithm for Markov Chain Monte Carlo can be endowed with certain "exploration steps" that periodically (at a given rate) allow it to explore the parameter space far from the current state to identify new, undiscovered modes. With this advancement, the MH algorithm can now effectively sample multi-modal distributions with a single chain! There's no need to launch multiple chains to enhance exploration. The proposed Intrepid MCMC is a general sampling method and can be applied in wide-ranging contexts from Bayesian inference to reliability analysis -- wherever MCMC is needed! Great work, Promit! #bayesian #MCMC #reliability #markovchain Johns Hopkins Whiting School of Engineering Johns Hopkins Department of Civil and Systems Engineering

88

2 Comments

A new paper with my favorite research partner, Matan Ben Zion! link: https://lnkd.in/dJT6sKez We looked at a seemingly very basic question - how does a drop of repulsive particles spread? It's a basic setting since almost all micro and nano-scale particles are, in fact, charged. Through theory, simulation, and experiment, we discovered that such a drop spreads very differently from a regular diffusive drop. First, surprisingly, it spreads more slowly. Second, the drop remains compact. In other words, unlike a drop of dye that when it spreads in water, its edges become blurred, here, the edges remain razor sharp even at long times. The theory we developed is very general and applies to all scales - from quantum particles in a superconductor to robots and collectives of animals or humans.

49

1 Comment