Amath SOW
New York (Open to Remote)| linkedin.com/in/amathsow | - |-
EXPERIENCE
Apziva
Remote
Artificial Intelligence Resident
December 2021- Present
● Used NLP techniques to rank potential talents identified on a job portal.
● Implemented machine learning algorithms on marketing data to forecast term deposits at banks.
● Built machine learning models to forecast customer satisfaction with 83% accuracy.
Intelligent Robot Learning Lab
Alberta/Canada
Reinforcement Learning researcher
June 2020 - January 2022
● Implemented four Human in Loop Reinforcement Learning(HCRL) algorithms using tensorflow framework :
TAMER (Training and Agent Manually via Evaluative Reinforcement), COACH (Convergent Actor Critic by
Human Feedback), DEEP-TAMER (TAMER using Deep neural network), DEEP-COACH (COACH using
Deep neural network)
● Built a new HCLR algorithm, Variance-Reduce COACH(VR-COACH) which interprets the human feedback
as reward and apply variance reduction technique on policy gradient commonly used in RL
● Experimented with VR-COACH in the classic MountainCar environment and DEEP VR-COACH in Goal
navigation task(Malmo Minecraft).
Stem_away
California/USA
Machine Learning Intern
June 2021-August 2021
● Learn selenium and Beautifulsoup to do web scraping
● Filtered and data scraped the “CarTalk” forum in order to recommend to car users the necessary information
needed to help navigate the forum
● Built a recommendation system using BERT pretrained model specifically for the “CarTalk” discourse forum
which improved user experience : link to the certificate
Huawei Technologies(3G/4G)
Senegal
Radio access network/Application and service engineer
January 2015- September 2018
● Involved in Radio Access Network (RAN) maintenance and monitoring for 3G and 4G.
● Involved in Application and Service (AS) maintenance for the network intelligent: CRM, CBS, API SMS,
USSD, Network KPI monitoring.
ORANGE(4G)
Senegal
IRadio access network Intern for 4G network
March 2014- June 2014
● Implemented IPSec to secure S1 and X2 interfaces for the Orange 4G network.
● Used tunneling with VPN, RSA asymmetric chiffrement with public key generation. Certification
Authority (CA) and Deployment on CENTOS Server.
PROJECTS
Predict if the customer will subscribe to term deposit marketing
●
●
●
●
●
●
Explored the data in order to know the distribution of the data
Applied imputation technique for unknown value on data(mean, median and mode)
Applied SMOTE technique by oversampling from the minority class(data highly unbalanced)
Determined analytically the segment of customers that are more likely to buy the investment product.
Built a stack of algorithms classifiers and train them using 5-fold cross validation. Decision Tree Classifier
gives us the best accuracy(89%) with a recall of 90% of class 1 and 70% of class 0.
Plotted the features importances about which features are going to be more useful for the next survey.
Predict if the customer is happy or not based on the answers they give to questions asked
●
●
●
●
Explored and cleaned the data (126 data samples)
Applied dummy encoding for categorical data
Used the chisq test(chi2) for feature selection in order to know which features are more correlated with target
variables.
Built a stack of algorithms classifiers and train them using 5-fold cross validation. XGBoost gives us the best
accuracy 78.12% which is greater than the expected accuracy(73%).
Variance Reduced Convergent Actor Critic by Human Feedback(VR-COACH)
●
●
●
●
●
●
●
Reproduced COACH algorithm using tensorflow
Built Episodic COACH by modifying COACH algorithm and computed the policy gradient for each timestep
but the gradient update is done at the end of the episode.
Observed that the gradient update is similar to the one in REINFORCE except that we have the human
feedback in place of the reward and the eligibility decay 𝜆 in place of the discount factor.
Built a new algorithm VR-COACH, which is an actor-critic based algorithm in which the human feedback is
interpreted as reward.
Experimented with VR-COACH in the classic MountainCar environment and demonstrated that it learns
faster than COACH and TAMER.
Upgraded our original VR-COACH to his DEEP version (DEEP VR-COACH) where agent’s policy is
represented as a deep neural network and apply Convolutional Auto Encoder strategy, a Feedback Replay
Buffer and entropy regularization in order to learn complex task in a reasonable time
Demonstrated the effectiveness of DEEP VR-COACH in the rich Malmo Mine-craft environment while
comparing it with Deep COACH and DEEP TAMER.
Cassava Disease Classification using the images of cassava and Computer Vision approaches and
semi-supervised learning:
●
●
●
●
●
Explored and transformed the data using fast.ai deep learning library
Used pretrained architecture(densenet101) and built our learner while using accuracy as metric and adam
optimizer.
Unfreezed the saved model, applied fine-tuning and slicing learning rate.
Used our pretrained model to predict labels for our extra images (semi supervised learning).
Trained the model on the new dataset on 30 epochs and end up with 92% of accuracy
EDUCATION
African Master in Machine Intelligence(AMMI)
Master in Machine Intelligence
Gaston Berger University
Electronic and Telecommunication engineer
Gaston Berger University
Bachelor in Maths, Physics and computer science
Ghana
March 2021
Senegal
November 2014
Senegal
August 2011
SKILLS & INTERESTS
Skills: Python (Scikit-Learn, PySpark), SQL, MongoDB, Doker, GCP,Jenkins ,Excel, Matplotlib, Seaborn, Linear
Regression, Logistic Regression, Classification, Neural Networks, Natural Language Processing(Word
embedding, LSTM, transformer, GRU, machine translation), computer vision(CNN, CAE, segmentation, object
detection, object tracking), reinforcement learning(MDP, Q-learning, DQN, Actor-Critic, HCRL), Keras,
TensorFlow, pytorch, Agile Methodologies
Interests: watching futuristic movies - discuss with friends and family - play soccer
