Meghana Bhange
This work adds two common approaches: fine-tuning large transformer models and sample efficient methods like ULMFiT. Prior work demonstrates the efficacy of classical ML methods for polarity detection. Fine-tuned general-purpose language representation models, such as those of the BERT family are benchmarked along with classical machine learning and ensemble methods. We show that NB-SVM beats RoBERTa by 6.2% (relative) F1. The best performing model is a majority-vote ensemble which achieves an F1 of 0.707.
This article presents the prototype design and testing of a long-range, self-powered IoT devices for use in precision agriculture and aquaponics. The devices are designed using the ultra-low power nRF52840 microcontroller with Bluetooth 5 support and ambient energy harvesting. A power of 942μW is harvested in an indoor environment. The devices are therefore suitable for both indoor and outdoor use, as natural sunlight will provide far more energy compared to artificial indoor lights. A line-of-sight range of up to 1.8km is achieved with the use of coded transmissions. However, the coverage area and range can be extended significantly by deploying the devices in multi-hop network topology. The custom multi-hop protocol provides energy efficient communication from any device in a wireless sensor network to a gateway while consuming an average of 267μW with a transmission interval of 5 minutes. The sensor data is transmitted to a gateway, which then forwards it to a local server or cloud service, where the data can be analyzed to optimize the production in agriculture and aquaponics.
This article presents the design and prototype implementation of a low-cost and short-range self-powered wireless IoT device based on energy harvesting for both indoor and outdoor applications. Prototyped devices are deployed in a star network configuration with a custom protocol. Based on measurements, devices achieve a line-of-sight range of 228.5m above 40m from the ground level. Nodes are powered based on energy harvesting from a small 0.36W solar panel and 120mAh lithium button cell as storage elements. The test in the well-lit room shows an average harvested power of 941.94μW over a period of 2.5 days, while under the low lighting conditions showed an average of 212μW over a period of 24h. From measurements, a fully charged rechargeable 120mAh cell lasts for 278 days with 55s transmission interval. Temperature, visible lights level and relative humidity sensors are integrated into the nodes.
I am currently researching model extraction attacks on Machine Learning systems with explanations APIs. I am modelling an adversary that can leverage the information provided by counterfactual explanations to build high-fidelity and high-accuracy model extraction attacks. I have benchmarked the model performance on Folktables dataset with extracted model gaining fidelity of around ~97.6%.
Hunters is autonomous threat hunting solution that leverages XDR and AI to detect hidden cyber threats.
Verloop.io is a conversational AI startup based out of Bangalore, India. Working on the homegrown intent recognition service using sentence-transformer to improve the performance and accuracy. Current F1 out-performs the previous by 1.4x Worked on Person-Name Detection system and upgrading it to spaCy v3 Upgraded Legacy service in Django with integrating the tests with CI/CD solving for tech-debt and improving the development efficiency Prototyped GPT3 for data generation, data annotation, intent recognition and bench-marked it against the existing services. Using the data-annotation pipeline to fastrack the client-specific annotation process.
Created a Person-Name extractor customised for multilingual conversations. TweakedFlair(Facebook’s Natural Language Processing library)to work on chatbot specific use cases in English, Spanish and French Evaluated performance of various language models like ULMFiT, VAMPIRE for low-resource language contexts Deployed the developed multilingual name-extractor to production; Final model achieves 47% improvement in F1 compared to previously deployed FastText mode
Prototyped a machine learning powered market place for streamlining transactions between buyers and sellers that intelligently estimates negotiated prices by observing behaviour in the first few manual negotiations.
Chatbots are an upcoming and automated way employed by businesses to communicate with their clients. An important aspect of personalising this communication is to employ natural language rather than use text boxes with strict bounds. As part of this, it is important to extract named entities(to understand the customer’s name) from messages written in unstructured, natural language. This problem is called Named Entity Recognition.
Named Entity Recognition for Chats can face several issues like inputs consisting of- Inconsistent grammer Emojies and other unexpected unicode characters Typos and inconsistent capitalisation. This talk would focus on creating an efficient NER for chats by tweaking the current state of the art NER’s
This talk focuses on how to use GPT-2 to create your own lyrics generator and has three parts