Deepank Dixit

ML enthusiast and explorer. Graduated with Masters in AI from Indian Institute of Science, Bangalore while working as a full-time Network Security Engineer at Cisco. Total 2 years of exposure to Data Science and ML, and 5 years of experience in Network Security and Collaboration solutions within Cisco Zero Trust Security solutions. Technical Writer and Public Speaker at conferences Cisco Offensive Summit 2021, 2022, Cisco SecCon 2021, QCon Plus 2021.

• Analyzing Exoplanets through Probabilistic Machine Learning in AstroML (2023)

Utilized Bayesian Regression to model equilibrium temperature based on planet's insolation flux, a critical factor in determining exoplanet habitability. Employed the Chow-Liu algorithm for structure learning to identify conditional dependencies among various exoplanet parameters. This approach allowed us to better understand the complex relationships within the dataset. Developed a classification model to categorize exoplanets as Small, Medium, or Large in comparison to Earth's size using the radius parameter (pl_rade). The model's performance exceeded random guessing, achieving an accuracy of 0.47 for the three classes. Employed Gaussian Mixture Models (GMM) to cluster exoplanets based on their habitability-influencing features. Through PCA, we reduced dimensionality, allowing us to visualize clusters effectively. Addressed the challenge of missing data by experimenting with GaussianProcessRegressor for 'pl_rade' feature imputation and iterative imputation for other parameters, resulting in meaningful insights and improved model performance.

• Estimating Missing Values with Bayesian ML (2023)

Developed probabilistic ML models to predict missing temperature and humidity values for 5 sensor nodes located at different positions. Employed a Bayesian linear regression model to estimate the probability distribution of the target variables, utilizing the likelihood and prior distributions to predict mean and variance. We also Explored the efficacy of a Linear Regression model and Random Forest Regressor, both with and without Kriging, considering the spatial data of the nodes. RMSE based evaluation revealed that the Bayesian linear regression and Random Forest Regressor with Kriging achieved the best results. Challenges such as high training time, RMSE errors, and compatibility issues while experimenting with Bayesian Neural Network and Gaussian process models.

• Neural Image Caption Generator ( 2022)

Tools, Packages, APIs– Tensorflow/Keras, Python, NumPy |Domain– NLP, Computer Vision

Used Transfer Learning to encode images of Flickr8k Dataset through Inceptionv3 CNN module, which acted as an Encoder. Preprocessed, vectorized the text captions, and prepared the word embeddings through Keras Embedding layer. Fed these word vectors with encoded image representations as input to a TransformerDecoder with 6 Multi-Head Attention Layers to predict the output. Model was trained using “Sparse categorical cross-entropy” loss, “Adam” optimizer and metric “accuracy”. Reported about 50% test accuracy. (https://colab.research.google.com/drive/1n8HopL4C7cT8XB5VBktWaIw6VL3B6vcv?usp=sharing)

• Adversarial Attacks in ML (2022)

Tools, Packages, APIs– PyTorch, Python, NumPy |Domain– Computer Vision

Using PyTorch, built and trained an MNIST classifier using Convolutional Neural Network (CNN). Generated Adversarial examples using FGSM (Fast Gradient Sign Method) and demonstrated model's vulnerability to FGSM and One-pixel attacks. The test accuracy came down to 57.96% from original 98.59%. Performed Adversarial Training as the defence against these attacks, post which the prediction accuracy on adversarial examples improved to 98%. Plotted the relevant visualizations for demonstration. (https://jovian.ai/deepankdixit0804/fooling-an-mnist-classifier)

• Time Series Weather Forecasting (2022)

Tools, Packages, APIs– Tensorflow/Keras, Python, Pandas |Domain– Sequential Data (Timeseries)

Used 7 years' worth of data from Dataset jena_climate_2009_2016.csv to predict future temperature. Ran and compared different recurrent neural networks like RNN, LSTM, GRU with a baseline non-ML model and reported the results. (https://colab.research.google.com/drive/1QcFkxpikC5-NssjzttFuX3yavOWSJ5ui?usp=sharing)

• Data Science @Cisco Green Belt credential, First and runner-up of several ‘Capture The Flag' challenges
• Speaker at Cisco Offensive Summit 2021, 22, 23, Cisco SecCon 2021, and QCon Plus 2021, on Zero Trust security
• Recipient of Cisco Security Interstellar Pioneer (Security Ninja Black Belt)– recognized in Cisco for contributing to Security
• Mentoring 5+ new hires on Network Security - AAA team in APAC region at Cisco (2021-22)
• Trained team of 15+ professionals on Cisco WebEx Meeting Server in APAC and EMEA region at Cisco (2020)
• Wrote several TechZone articles on troubleshooting and deploying use cases in Network Access Control
• Tested and pointed defects in the upcoming ISE patches as rolled out on Cisco official release website.
• Cisco Innovation Challenge - WebEx bot framework to create WebEx space called Shuri to help TAC troubleshoot issues

Generative AI with Large Language Models