Reservoir computing is a promising framework for signal processing applications, but the optimization of such physical reservoirs remains an important challenge in the field. In this work we address this challenge with a new technique based on the use of a delayed input and we test it using an experimental optoelectronic reservoir. We demonstrate that this technique can replace the standard hyperparameters optimisation of reservoirs with a much simpler approach based on the scan of only two parameters. We test this approach on tasks of different nature, confirming its superiority in terms of ease of use and implementation to the standard hyperparameters scan.

Optical manipulation and tomographic imaging play critical roles in biomedical applications, however, applying these technologies to hard-to-reach regions remains challenging. We introduce a series of innovative AI-driven methods designed to facilitate both high-fidelity light field control and image reconstruction through a multicore fiber-optic system. Our approach enables precise, controlled rotation of human cancer cells around all three axes, enabling 3D tomographic reconstructions of these cells with isotropic resolution. The integration of these advanced optical and computational techniques culminates in a powerful optical fiber probe, capable of sophisticated optical manipulation and tomographic imaging, offering new perspectives for optical manipulation and its applications.

This study utilizes a novel method for learning representations from chest X-rays using a memory-driven transformerbased approach. The model is trained on a low-quality version of the MIMIC-CXR dataset, utilizing 17,783 chest X-rays that contain at most 3 views. The model uses a relational memory to record crucial information during the generation process and a memory-driven conditional layer normalization technique to integrate this memory into the transformer's decoder. The dataset is divided into distinct sets for training, validation, and testing. We aim to establish an intuitively comprehensible quantitative metric, through vectorization of the radiology report. This metric leverages the learned representations from our model to classify 14 unique lung pathologies. The F1-score measures classification accuracy, indicating the model's viability in diagnosing lung diseases. We also have introduced the use of Large Language Models (LLMs) for evaluation of the generated reports accuracy. The model's potential applications extend to more robust performance in radiology report generation.

In this research we explored the use of Mutual Connectivity Analysis with local models for classifying Autism Spectrum Disorder (ASD) within the ABIDE II dataset. The focus was on understanding brain region differences between individuals with ASD and healthy controls. We conducted a Multi-Voxel Pattern Analysis (MVPA), using a data-driven method to model non-linear dependencies between pairs of time series. This resulted in high-dimensional feature vectors representing the connectivity measures of the subjects, used for ASD classification. To reduce the dimensionality of the features, we used Kendall’s coefficient method, preparing the vectors for classification using a kernel-based SVM classifier. We compared our approach with methods based on crosscorrelation and Pearson correlation. The results are consistent with current literature, suggesting our method could be a useful tool in ASD research. Further studies are required to refine our method.

Over 75% of all FDA-cleared software as a medical device relate to use cases in radiology. Despite this large prevalence, the current status-quo of training artificial intelligence (AI) tools entails using unimodal (imaging-only) algorithms. Moreover, retraining such models for new tasks requires using training supervised AI algorithms from scratch, using manually curated labels from scratch, even if it may be for the same modality or anatomy. In radiology, generating such labels requires expensive clinical expert time, limiting the development of capable AI models across tasks. In this presentation, I will describe the development and use of multi-modal vision-language models (VLMs) for radiological applications. VLMs present numerous benefits such as zero-shot classification, label-efficient adaptation to varying tasks, and improved robustness. Such new capabilities provided by VLMs is poised to usher in a new era of models for solving current and future challenges in radiology.

Raman spectroscopy (RS) is a real-time, label-free, and non-invasive spectral sensing technique that can quantify the biochemical composition of biological tissues and other substances. However, Raman scattering is a weak effect and relies on long acquisition times across multiple acquisitions to produce a robust signal. Decreasing this collection time, as required in many time-sensitive in-vivo clinical applications, results in a signal with significant noise, which hinders interpretation. Various machine-learning (ML) denoising methods have been proposed for analyzing RS signals, but very few have successfully provided an accurate acquisitional denoising algorithm that works on a broad dataset across various real-life use cases. In this pilot project, we assess the feasibility of using convolutional neural networks (CNNs) and encoder-decoder transformer -based models for acquisitional spectral denoising. We utilize in vivo RS data from the human esophagus for testing our model to demonstrate its robustness on low signal-to-noise ratio spectra.

Current approaches to optical neural networks are only focused on the arithmetic operations, such as matrix vector multiplication and in some cases the activation function. The control flow to enable large matrices, multiple connected layers or methods to increase precision are completely delegated to digital electronics. Here we present our approach to enabling all-optical control flow and precision control to enable full optical AI inference in the future, eliminating the need to jump between the electronic and optical domain for each layer, to increase performance and efficiency.

Optical neuromorphic computing marks a breakthrough over traditional digital computing by offering energy-efficient, fast, and parallel processing solutions while challenges remain in incorporating nonlinearity efficiently. Leveraging nonlinear wave dynamics in optical fibers as a computational resource may provide a solution. Our research demonstrates how femtosecond pulse propagation in optical fibers can emulate neural network inference, utilizing the high phase sensitivity of broadband light for creating nonlinear input-output mappings akin to Extreme Learning Machines (ELMs). Experimental results show high classification accuracies and low RMS errors in function regression, all at pico-joule pulse energy. This indicates our method's potential to lower energy consumption for inference tasks, complementing existing spatial-mode systems. We also investigated femtosecond pulses' nonlinear broadening effects – self-phase modulation and coherent soliton fission – demonstrating their distinct impacts on classification tasks and showcasing broadband frequency generation as a powerful, energy-efficient tool for next-generation computing.

The rapidly emerging Generative AI technology that generates artificial images of real objects demands the development of a Detective AI technology to offer an efficient solution to solve the problem of distinguishing real and AI-generated images with high accuracy. The Generative AI, such as the Generative Adversarial Networks (e.g., BigGAN), can generate images that in turn can trick the human visual system and confuse human intelligence from seeing the true differences between AI-generated and real images. The Generative AI models, while capable of generating realistic images by focusing on image features, they lack efficiency in capturing complex optical features, including reflection, refraction, and shadows, that can leave traces and clues in the 8 bit-planes of a grayscale image. This paper proposes an approach that utilizes topological properties of the bit-plane images. It detects image components and characterizes their connectivity and adjacency relationships. The feature vectors that are generated using cross-correlation of these connected components in the bit-planes are used. A feature space that is constructed using these feature vectors is utilized to train and develop random forest (RF) classifiers for classifying AI-generated and real images. In a simulation with 200 BigGAN-generated images and 248 real images of house-finch birds, a random forest classifier is developed and validated. The results show, with a careful tuning of the RF parameters, the BigGAN-generated images and their real images can be classified with the F1-score and precision-score of about 81% and 82%, respectively. This research suggests that the cross-correlation features of the connected components can help distinguish real and AI-generated images.

We developed a computational model to simulate contours of entangled lambda DNA. These simulations were used to generate super-resolution DNA images for training a deep neural network (ANNA-PALM) to reconstruct DNA contours from localization images. Our approach enabled reliable contour prediction from microscopy images captured at fast time scale. Analysis of experimental data revealed bright and dark DNA segments, potentially linked to local microviscosity effects imposed by entanglement loci. Our integrated computational modeling and deep learning workflow can provide mapping of topological constraints on polymer motion in diverse materials.

In the current artificial intelligence (AI) framework, the explainability of AI is buried under the black-box nature of the implicit implementation of the latent feature space of an AI architecture. However, the explainability of AI can be enhanced by explicitly defining its feature space, quantifying the similarity (or dissimilarity) and orthogonality (or non-orthogonality) properties between its feature vectors, and extracting common features (projection onto subspaces) of its feature vectors. Hence, this paper presents an approach that defines a theoretically infinite family of features space (tIFFS) that uniquely combines the distinctive properties of inner product and orthogonality operations between feature vectors, and the projection of feature vectors onto subspaces in a Hilbert space. The tIFFS approach utilizes the concept of “infinite mixture model” of the (i) Bayesian Gaussian mixture model (also called the Dirichlet Process mixture model (DPMM)), (ii) spectral density of the DPMM output in the Fourier domain, and (iii) Hilbert space that is formed by infinite dimension function space (IDFS). The phase information that carries the orthogonality, orientation, and gradient features of the DPMM output in the Fourier domain is also utilized to precisely define subspace boundaries and capture similarity and dissimilarity features of the feature vectors. In addition, the tIFFS approach adapts principal component analysis to integrate the well-defined orthogonality properties of eigenvectors and eigenvalues of the covariance matrix of the feature vectors in Hilbert space. A simulation is conducted to develop random forest (RF) classifiers to classify backyard birds using tIFFS. The simulation with images of 147 Northern Cardinals, 147 American Robin, and 147 House Finches show the RF-classifiers that achieve the precision-score of about 88% and F1-score of about 86% can be developed by fine tuning the model and eigenvector (or eigenvalue) parameters. Hence, it shows that the tIFFS can capture suitable inner product, orthogonality, and projection properties in the IDFS Hilbert space.

This paper proposes a unique concept of colony of artificial intelligence (AI). A colony of AI is defined as a family of AI-agents that mimic the behavior of a biological system. The natural phenomena of biological systems, including colony of ants and colony of humans, is based on the idea that genetic evolution can occur through biological reproduction. Therefore, this paper defines a terminology, “marriage of AI-agents” to allow marriage between two AI-agents to produce unique offspring. It adapts the theory of genetic algorithm and utilizes the crossover and mutation techniques to build a colony of AI. AI models generally consists of a pair of parameters (kernel weight and bias) that they learned from an environment. These parameters are considered AI-genetic information for a colony of AI. A marriage allows the exchange of this AI-genetic information between two AI-agents through a crossover technique to produce a child AI-agent. Mutation is also used to make minor random changes to the child’s AI-genetic information to make it a unique AI-agent. One of the uniqueness of the proposed approach is the bias randomization of a parent AI-agent using Gaussian or uniform distribution (before applying the crossover technique) so that the flexibility to adjust classification boundaries is enhanced. This approach recommends the switching (crossover) of 50% of AI-genetic information from each parent, while modifying the bias parameter of one of the parents. This unique crossover-mutation technique with a modified AI-genetic information allows the parent AI-agents to produce offspring with an increased performance. Simulations are conducted for building a colony of AI using a pretrained VGG16 model and the CIFAR-10 dataset. Current simulations show that the flexibility of a parent (not of both parents) improves the performance of the child AI-agent by 6% (from 72% to 78%) on average with 10 epochs.