AKSHEIT SAXENA

Context Aware Crowd Management With Stakeholder Guidance (Crowd Vision / Computer Vision)

• Developed a general-purpose holistic system for resource-constrained environments to aid CCTV-based crowd management tasks under stakeholder guidance with contextual awareness for Cisco's Smart Camera solutions.
• Utilized state-of-the art Point Query Quadtree transformer for crowd-counting and utilized stakeholder guidance for adaptive crowd capture
• Created an ensemble classification model for crowd context classification based on in-real time crowd-density.
• Create maximum occupancy prediction model for chaos prevention
• Enhanced scene-based understanding using predefined mapping.

Denoising Diffusion Probabilistic Model and Momentum Contrast for Unsupervised Visual Representation Learning Implementation (Generative AI)

• Constructed a DDPM on the animal face data. Experimented with different values of backward inference time by plotting a grid of 10 by 10 images for all three cases and compute the FID. Used VAE to implement a class-conditional DDPM on its latent space. Used classifier guidance with the score-based formulation.
• For MoCO, implemented Self-supervised learning using Momentum Contrast Encoder method (MoCO). Tried 3 sets of augmentation methods. Implemented a linear classifier on the learnt representations. Compared with a full-blown CNN, that has the same number of parameters. Recorded the reduction in the amount of supervised data needed.

GMM, Vanilla-VAE and Beta-VAE Implementation (Generative AI)

• Constructed a GMM by implementing the EM algorithm, on all the images in the dataset, by downsizing the data to 28 by 28 pixels using diagonal Covariance. and using EM iteration progresses. Post training, generated and plotted 100 images in a grid of 10 by 10 with all three GMMs separately.
• Implemented a vanilla VAE with MSE for conditional likelihood. Plotted 10 by 10 grids of both reconstructions and generations. Plotted the loss curves for likelihood, KL, and the combined terms.
• Implemented a beta-VAE with 4 different values of beta and plotted 10 by 10 grids of generated and reconstructed data. Documented the observation in terms of change in results with varying beta.

Implementation of Deep Convolutional GAN and Wasserstein's GAN (Generative AI ) 

• Trained a DC GAN on the animal dataset with the usual GAN loss. Plotted the loss curves for the Generator and Discriminator networks along with 10 by 10 grid of images for generated images. Experimented with the number of times the generator and discriminator are trained and documented the changes in the behavior.

Depression Detection using Speech and Text Analysis

• Developed a deep learning model for detection of depression by analyzing speech patterns and textual data.
• Employed acoustic feature extraction for speech and sentiment analysis for text to capture relevant features that areindicators of depressive symptoms.
• Built and trained models including Support Vector Machines (SVM), Convolutional Neural Networks (CNN), LSTM,BiLSTM and Transformers to classify depression levels based on the extracted features.
• Validated on test data to ensure the accuracy and reliability of the trained depression detection model.

Crowd Segmentation and Counting using Few Shot Learning

• Used a Few-Shot Learning approach to tackle the fundamental problem of crowd counting in densely crowded scenes.
• Utilized meta-learning in the context of few-shot learning to enhance model adaptation to target crowd scenes with only a few training samples. Thus overcame the challenges related to the limited availability of annotated data and the constraintsof conventional supervised learning methods that depend on the requirement of vast amounts of labeled training data.
• Trained the model for rapid adaptation to target crowd scenes with minimal examples during testing, ensuring it generalizeswell to new scenes and performs well on the task of crowd counting.

Community Detection and Link Prediction in Social Networks

• Performed community detection on a Facebook dataset by implementing the Louvain algorithm from scratch and applied spectral clustering to analyse the structural properties and relationships within the detected communities.
• Implemented graph machine learning techniques like GNNs for predicting missing links in the network.
• Used the Metropolis-Hastings sampling technique to sample positive and negative samples to train the model effectively.
• Incorporated graph measure features to enrich the feature set for link prediction and was able to get a ROC-AUC score of0.96 for link prediction between nodes.

Ethosight: A Reasoning-Guided Iterative Learning System for Nuanced Perception based on JointEmbedding & Contextual Label Affinity

• Developed a cutting-edge open-source project that introduces a zero-shot video analytics system, which is both flexible and adaptable. What sets Ethosight apart is its ability to start with no pre-existing trained data, relying solely on 14 ground truth labels.
• This system is uniquely user-friendly, as it can be specified through natural language or simple keywords. The power of Ethosight lies in its use of joint embedding models and reasoning mechanisms, which are enhanced by ontologies such as WordNet and ConceptNet. WordNet organizes words into synonym sets, aiding in tasks like word sense disambiguation and text analysis, while ConceptNet represents general knowledge about the world, which is invaluable for natural language understanding and commonsense reasoning

• Active member and former Vice President at Toastmasters International
• Served as IT committee Chief Panelist at multiple occasions at Ghaziabad Management Association (GMA) serving local industries and business with technological solutions.

• Won several district level Toastmasters International Speech Competitions.
• Honored as Chief Panelist at several Indian Management Association Technology (AI & Cybersecurity) Awareness sessions

Technology

• Oracle Cloud Infrastructure 2024 Generative AI Professional
• Ethics In the Age of AI (LinkedIn)
• Neural Networks for Machine Learning (Coursera)
• Cisco Certified Network Professional (CCNP)
• Cisco Certified DevNet Associate
• Cisco Certified Network Associate Routing and Switching (CCNA)
• Hardware Security (Coursera)

Language

• German- A1 Proficiency Level (Goethe-Institut e.V.)

Business & Processes

• ITIL IT Service Management

Latapie, Hugo et al. “Ethosight: A Reasoning-Guided Iterative Learning System for Nuanced Perception based on Joint- Embedding & Contextual Label Affinity.” (2023).